GPT-5.6, OpenAI's new model that thinks faster than the competition and costs less

What is GPT-5.6 and why is it three models at once

OpenAI has officially opened access to the GPT-5.6 family following a limited preview launch. It's not one model, but three, each built for a different type of task. Sol is the flagship, the most powerful version. Terra is a balanced model for everyday work. And Luna is the fastest and most affordable of the three.

The core idea behind this launch is simple, OpenAI wants stronger results to stop automatically meaning a higher price tag. The model was trained to use fewer tokens, meaning fewer "units of text", to complete the same work. The result, GPT-5.6 delivers equal or better quality while costing less to run than both previous OpenAI models and competing models.

Worth mentioning separately is a new mode called ultra. It runs several AI agents in parallel on one complex task to get results faster. More on that below.

Why GPT-5.6 runs faster and cheaper than before

On a test called Agents' Last Exam, which checks how well an AI handles long professional tasks across 55 different fields, GPT-5.6 Sol scores a record 53.6 points. That's 13.1 points ahead of Claude Fable 5, one of its main rival models. Even in medium reasoning mode, GPT-5.6 beats Fable 5 by 11.4 points while spending roughly four times less money on the task.

A similar pattern shows up in the smaller models. Terra and Luna outperform Fable 5 while costing around sixteen times less. On another major benchmark that measures a model's overall intelligence across many areas at once, GPT-5.6 Sol comes within one point of Fable 5, while completing tasks 61 percent faster and at roughly half the cost.

The takeaway here, GPT-5.6's main advantage isn't that it got smarter "at any cost", it's that it learned to deliver comparable or stronger results while spending noticeably less. In practice, that means you can either get more work done for the same money, or get the same amount of work done for a lot less.

What's new in coding and running multiple AI agents at once

GPT-5.6 Sol is OpenAI's best coding model to date. On the Artificial Analysis Coding Agent Index, a benchmark that evaluates an AI's performance as a developer, the model scores a record 80 points, beating Fable 5 by 2.8 points. It does this while using less than half the output tokens, taking less than half the time, and costing roughly a third less. The model also sets new best results on tests measuring command-line work and complex engineering tasks in real codebases.

A separate new feature is called programmatic tool calling. Previously, a model had to keep "checking in" with itself at every step, passing results back and forth. Now GPT-5.6 can write a small program that coordinates the needed tools on its own, filters out unnecessary data, and keeps working without extra intermediate steps. That makes complex tasks faster and cheaper.

For the most demanding tasks, there's an ultra mode. It runs several AI agents in parallel, four by default, and up to sixteen in some tests. Each agent works on its own part of the task at the same time as the others. According to tests on web browsing, cybersecurity, and terminal work, this parallel approach delivers noticeably stronger results in less time, though at the cost of higher token usage.

The simple takeaway here, standard mode is plenty for everyday tasks. Ultra is worth turning on only when the task is genuinely complex and speed matters more than saving on tokens.

Design, documents, and everyday office work

GPT-5.6 made a real leap forward in design. Given only a high-level description of the task, the model creates clean, ergonomic, functional interfaces. Thanks to stronger computer-use skills, it no longer just writes the underlying code, it actually checks how the finished result looks and makes final touch-ups before handing the work back.

A similar improvement shows up in everyday office work. The model takes scattered information from documents, chats, and tools like Slack, Notion, Microsoft 365, and Google Drive, and turns it into a polished, ready-to-use output. On BrowseComp, a test measuring web research and information handling, GPT-5.6 Sol scores a record 92.2 percent, and on OSWorld 2.0, a computer-control test, it hits 62.6 percent, surpassing the Opus 4.8 model while using 85 percent fewer output tokens.

The model also got better at building presentations, documents, and spreadsheets. It can create fully editable presentations from scratch, and when given a company template or reference deck, it accurately matches the style, fonts, and structure for new slides. It handles spreadsheets and financial calculations more precisely too.

Cybersecurity, science, and accelerating research inside OpenAI

GPT-5.6 is OpenAI's strongest model yet in cybersecurity. On a test that checks whether an AI can take a code vulnerability all the way to a full attack, the model scores 73.5 percent versus 47.9 percent for the previous GPT-5.5. On another test that asks the AI to turn a real vulnerability into a working exploit, the result nearly doubled, from 15.1 to 24.9 percent within a two-hour window, reaching 33.7 percent given six hours.

It's important to understand that these capabilities cut both ways. The same skills that help an attacker find a vulnerability help defenders find and patch that same vulnerability first. That's why OpenAI reserves access to the model's strongest defensive capabilities for verified users and organizations through a dedicated program.

Inside OpenAI itself, researchers already use GPT-5.6 to speed up their own work, from finding bugs to analyzing experiments. Over the past six months, the share of computing power spent on internal AI-assisted coding grew a hundredfold, and internal use of AI agents grew roughly twenty-two times over.

How much does GPT-5.6 cost and where can you try it

GPT-5.6 is already available in ChatGPT, Codex, and through the company's API, with the full worldwide rollout completing within a day. Plus, Pro, Business, and Enterprise users get access to GPT-5.6 Sol, while Pro and Enterprise users can also choose the enhanced Sol Pro version for the hardest tasks. Free users get access to the Terra model.

Developer pricing is calculated per million tokens and breaks down like this,

• Sol, the flagship model, 5 dollars for input tokens and 30 dollars for output tokens

• Terra, the balanced model, 2.5 dollars for input and 15 dollars for output

• Luna, the most affordable model, 1 dollar for input and 6 dollars for output

That price spread makes the choice simple, Sol fits genuinely complex tasks, Terra covers everyday work just fine, and Luna is the cheapest and fastest option for simple, high-volume requests.

GPT-5.6 shows that the AI race is no longer just about the strongest result at any cost, it's about who can deliver a strong result faster and cheaper. OpenAI released three models built for different tasks and budgets, and the price gap between them and the competition turned out to be substantial.

But that raises another question, how do you quickly test Sol, Terra, and Luna, and compare them against other AI models on the market, without setting up a separate OpenAI subscription and dealing with dollar payments directly. Juggling several paid accounts just to run a comparison isn't exactly convenient.

That's exactly what unitool.ai is for. Get one subscription and unlock access to the newest models, including the entire GPT-5.6 lineup, no foreign card needed and no separate plans to figure out for each platform. See for yourself which model and which mode actually fits your task best.